A score of 90 is considered passing: How exactly to build and evaluate a high-quality dataset

Let's start with a fact that many people are reluctant to face.

Chances are that 90% of the hundreds of terabytes of data lying in your company's servers are "garbage" — not the kind that can't be used, but the kind that hasn't undergone professional processing and doesn't meet the standards for AI training.

In September 2025, the National Data Bureau released 104 typical cases of high-quality datasets, covering 11 major fields such as scientific research, industrial manufacturing, healthcare, and financial services. What does this mean? Out of the tens of thousands of datasets across the country, only 104 were recognized as "typical."

The proportion is less than 1%.

On July 3, 2026, Liu Liehong, the director of the National Data Bureau, announced at the Data Element Development Forum of the Global Digital Economy Conference that the country will accelerate the introduction of national standards for the format requirements, quality evaluation, and data annotation of high-quality datasets.

The standards are coming. This means that in the future, "high-quality" will no longer be a vague adjective, but a hard and quantifiable indicator with a clear passing score.

What's the passing score? 90 points. On a 100-point scale, a dataset must score over 90 points in all three dimensions to be considered a "high-quality dataset."

This article aims to clarify two things: how to build and how to evaluate high-quality datasets.

I. First, understand: What is a "high-quality dataset"

On June 3, 2026, the National Data Bureau issued the "Implementation Plan for Promoting the Construction of High - quality Datasets in the Industry," which gave a clear definition:

A high - quality industry dataset is a collection of industry data that has undergone data processing such as collection and processing and can be directly used for the development and training of artificial intelligence models, effectively improving the model's performance. It includes general industry and specialized industry datasets.

Pay attention to three key words:

"Processed" — It's not enough to just pile up raw data. It must go through professional processing such as collection, cleaning, processing, and annotation.

"Directly usable" — You can use it to train models right away without a lot of pre - processing work. This is what the industry calls "AI - Ready."

"Effectively improve" — Using your dataset genuinely improves the model's performance. It's not a waste of effort but has tangible results.

This clearly differentiates high - quality datasets from ordinary data resources. The raw logs on your hard drive, unprocessed user behavior data, and messy Excel spreadsheets — these are "data resources," not "high - quality datasets."

The relationship between them is like that between iron ore and steel. Iron ore is a resource, and steel is a processed product. AI models need steel, not iron ore.

II. How to build: Six major actions and the "1+1" approach

The implementation plan of the National Data Bureau deploys six special actions, which form a complete chain for the construction of high - quality datasets:

1. Strengthen the foundation and expand capacity — Solve the problem of "availability"

Focus on 19 key fields (such as scientific research, industrial manufacturing, agriculture and rural areas, smart energy, transportation, financial services, and healthcare) and 5 innovative fields (low - altitude economy, embodied intelligence, intelligent driving, smart ocean, and biological manufacturing) to accelerate the construction of high - quality industry datasets.

The approach is clear: First, conduct an inventory — establish a list of data resources and a list of dataset requirements; then, find leading enterprises in the industry chain to take the lead — support leading enterprises to promote the collaborative construction of the upstream and downstream of the industrial chain in the form of a consortium; finally, adopt a multi - modal approach — cover text, code, images, audio, video, point clouds, time - series data, and scientific data comprehensively.

2. Tackle data annotation — Solve the problem of "comprehensibility"

Data annotation is not just about labeling. It is a process of injecting knowledge and experience into training data.

The plan proposes an important transformation: from a "human - dominated" approach to a multi - level annotation model of "human - machine collaboration with in - depth expert participation."

Specifically, develop three types of intelligent annotation services:

• "Model pre - annotation + manual calibration"

— Let the AI do the initial annotation, and then humans check and correct it.

• "Manual annotation + model verification"

— Humans do the annotation first, and then the AI conducts quality control.

• "Model pre - annotation + model verification"

— The AI does both the annotation and verification, and humans only handle disputed items.

Even more important is expert - led annotation. Establish a certification mechanism for industry experts to involve real - world industry experts in the professional knowledge annotation of instruction fine - tuning and reinforcement learning stages. Some professional knowledge can't be annotated by ordinary annotators — medical images need radiologists, legal documents need practicing lawyers, and industrial quality inspections need senior engineers.

3. Improve quality and efficiency — Solve the problem of "goodness"

This step is the core. The plan proposes four quality standards: structural integrity, content diversity, annotation accuracy, and model adaptability.

In terms of technical means, encourage the use of data intelligent filtering and matching technologies to build high - knowledge - density datasets that are "more refined and more powerful" and reduce the cost of training and inference. At the same time, give full play to the positive role of data synthesis — use models and simulation systems to generate data to solve the problem of constructing datasets for scarce scenarios.

4. Empower applications — Solve the problem of "usability"

A dataset that is built but not used is just a decoration in the digital warehouse. The plan proposes to create a closed - loop "data flywheel" application:

Scenarios drive data → Data drives models → Models empower applications → Applications create value → Generate new data → Return to data supply

Once this closed - loop starts running, the dataset will become better and richer with more use.

5. Management and service — Solve the problem of "manageability"

Build a management system that covers the entire lifecycle of collection, cleaning, processing, annotation, quality control, evaluation, iteration, and auditing. Construct a national dataset management and service system that is "physically distributed but logically centralized" to achieve interconnection of dataset catalogs and supply - demand information.

6. Unlock value — Solve the problem of "worthiness"

Explore the path of dataset assetization: registration, evaluation, pledge financing, capital contribution in the form of shares, asset securitization, data trust, and data insurance. Foster a market consensus of "paying for data" — promote the inclusion of data procurement in budget preparation and initiate data procurement practices in government departments, state - owned enterprises, and model enterprises first.

III. The three transformations mentioned by Liu Liehong are more important than you think

At the forum on July 3, Liu Liehong made a very crucial statement, revealing a fundamental shift in the thinking of high - quality dataset construction.

Transformation 1: From "emphasizing format uniformity" to "semantic consistency"

In the past, people thought that for data to be circulated and used to train AI, the formats had to be unified first. It was too chaotic with some storing data in JSON, some in XML, and others in Excel, so unification was necessary.

But Liu Liehong said:

If all departments, industries, and enterprises are required to unify the underlying data storage formats, the cost will be extremely high, and the efficiency will be low. In many fields and scenarios, it is not practically feasible.

The new thinking is that what is really needed is an interoperable rules layer, rather than a unified underlying storage format.

For example, if you speak French, I speak Chinese, and he speaks Japanese, there's no need to force everyone to speak French. What's really needed are translation rules — as long as the semantics can be understood, the formats can be different.

Transformation 2: From "static alignment" to "dynamic adaptation"

In the past, it was "one - time cleaning and alignment" — spending three months to unify the data formats and then using them all the time. However, data is dynamic, business is changing, and data sources are changing. Static alignment quickly becomes outdated.

The new approach is "dynamic adaptation" — relying on mechanisms such as metadata mounting and semantic mapping to achieve data with built - in standardized semantics to support automatic context recognition by machines.

What does this mean? Each piece of data comes with its own "instruction manual." When the machine reads the data, it can automatically understand the data's meaning, format, source, and purpose. There's no need for manual alignment; the machine can handle it on its own.

Transformation 3: From "classified measures" to "national standards"

Liu Liehong clearly stated that in the next step, a strategy of classified measures and hierarchical promotion will be adopted. At the national level, unified standards in many aspects such as object standards, metadata semantics, and interface protocols will be studied and formulated, and national standards for the format requirements, quality evaluation, and data annotation of high - quality datasets will be accelerated.

This means that in the future, there will be national standards for "high - quality." It's not up to individuals to claim that a dataset is high - quality; it has to be constructed, evaluated, and certified according to national standards.

IV. How to evaluate: A three - dimensional evaluation system with 90 points as the passing score

This is the most practical part. The "Quality Evaluation Specification for High - Quality Datasets" issued by the National Technical Committee for Standardization of Information Technology (TC609) has built a three - dimensional evaluation system:

Dimension 1: Documentation (Data passport)

A dataset must be accompanied by a complete "data passport," which includes four parts:

• Basic information

— Dataset name, version, scale, creation time, and update frequency

• Content characteristics

— Data type, coverage scope, subject area, and language

• Construction process

— Collection method, processing flow, annotation specification, and quality control measures

• Application instructions

— Applicable scenarios, usage limitations, known limitations, and citation methods

This dimension examines "transparency" — enabling users to quickly determine whether the dataset is suitable after obtaining it, rather than treating it as a mystery.

Dimension 2: Data quality (Eight core indicators)

This is the most critical part. The eight core indicators include:

• Format compliance

— Whether the data format complies with standard specifications

• Content authenticity

— Whether the data can be traced and is not fabricated

• Completeness

— Whether key fields are missing and whether the coverage is sufficient

• Accuracy

— Whether the data values are consistent with the real situation

• Consistency

— Whether the information of the same entity in different records is contradictory

• Timeliness

— Whether the data is within the valid period and whether it is updated in a timely manner

• Security

— Whether it contains sensitive information and whether it meets data security requirements

• Diversity

— Whether the data distribution is balanced and whether it covers enough scenarios

Pay attention to the "content authenticity" indicator — it emphasizes data traceability. This means that there must be a complete record of where your data comes from and what processing it has undergone. You can't claim that a batch of data with unknown origins is "high - quality."

Dimension 3: Model application (Five indicators)

This dimension focuses on "practical value" — whether the dataset can really improve the model's performance. Five indicators verify the actual contribution of the dataset to model training.

The most crucial one is model adaptability — it requires empirical testing to prove that the performance of the target model has indeed improved after using your dataset.

This directly breaks the "quantity - only theory." It's not that the more data, the better. It depends on whether the data is useful for the model. 10,000 high - quality annotated data points may have a greater impact on model improvement than 1 million low - quality data points.

Scoring rules: On a 100 - point scale, score over 90 points in all three dimensions

Each of the three dimensions is scored on a 100 - point scale. A dataset must score over 90 points in all three dimensions to be recognized as a "high - quality dataset."

The certification process is divided into four stages: application → preliminary evaluation → re - evaluation → certification. It's a strict process, not just a formality.

V. How much is a high - quality dataset worth?

This is not just empty talk. The market has already answered this question with real actions.

According to industry research, the premium for high - quality datasets can reach 20% - 30%. For datasets of the same scale, those certified as high - quality are 20% - 30% more expensive than uncertified ones.

This is driving the formation of a market mechanism of "pricing by quality" — in the past, data transactions were "priced by quantity" (a certain number of data points for a certain price), but in the future, it will be "priced by quality" (the higher the quality score, the higher the price).

The business model is also evolving. The plan encourages the development of three service forms:

• "Subscription model"

— Pay monthly/annually to continuously obtain updated data

• "Mall model"

— List datasets on the data trading platform for self - service selection

• "Customization model"

— Customize datasets according to needs with dedicated services

A more advanced exploration is "token trading" — building a quantifiable and priceable data value system based on tokens. The value of your data depends on how many tokens are called by the model.

The path to assetization is also being cleared. Datasets can be registered, evaluated, used for pledge financing, capital contribution in the form of shares, or even securitized. That is to say, the high - quality datasets you build can not only be sold for money but also used as assets for bank mortgages.

VI. What should enterprises do now?

After all this, if you're the head of an enterprise, what should you do now? Here are five suggestions:

First, align with national standards and prepare in advance.

Liu Liehong has clearly stated that national standards for the format requirements, quality evaluation, and data annotation of high-quality datasets are about to be introduced. Don't wait for the standards to come out before taking action. Start sorting out your datasets now according to the framework of the three - dimensional evaluation system: Does your dataset have a "data passport"? What score can it get on the eight - item quality indicators? Has model adaptability verification been conducted?

Second, work backwards from scenarios and don't blindly accumulate data.

The common feature of the 104 typical cases is "scenario - driven." First, figure out what problems your dataset is supposed to solve —